Research map / Video / Video generation
Q-ARVD: Quantizing Autoregressive Video Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Q-ARVD, a post-training quantization framework for autoregressive video diffusion models that weights calibration by frame-wise final-quality sensitivity and uses outlier-aware adaptive dual-scale quantization.
Problem
Quantizing autoregressive video diffusion models (ARVDs) is largely unexplored; existing diffusion/LLM quantization methods perform suboptimally because of unbalanced frame-wise sensitivity from error accumulation and heterogeneous weight channel outliers.
Approach
Measures each frame's quantization sensitivity by quantizing only that frame and observing the change in the whole generated video, uses these scores as reconstruction loss weights, and splits outlier input channels (detected via Modified Z-score with a minimum-magnitude constraint) into a separate quantizer per layer.
Key results
- On self-forcing W4A8, Q-ARVD reaches FVD-FP 116.26 and LPIPS-FP 0.466, versus 121.83-156.70 for variants and 124.20-260.14 for baselines' FVD-FP (best baseline PTQ4DiT 124.20).
- On causal-forcing W4A6, Q-ARVD achieves FVD-FP 140.38 vs 210.28 for SVDQuant and 375.82 for MinMax (lowest bit-width setting).
- INT8 (W8A8) deployment yields 1.30x latency speedup (18.02s to 13.85s on NVIDIA A6000) and 1.97x model size reduction (2.64GB to 1.34GB).
Surprising findings
- VBench scores show limited discriminability for quantization; W4A6 Q-ARVD scores higher than W8A8, W4A8 and even BF16 on some VBench metrics.
- Heuristic exponential frame weighting underperforms uniform weighting in some settings, and reversed weighting underperforms uniform weighting.
Limitations
- Frame sensitivity is exploited only during reconstruction, not for mixed-precision quantization.
- Deployment uses Triton kernels; hand-written CUDA kernels could give further efficiency gains.
- Evaluation relies on reference-based metrics against the FP model, with VBench judged less discriminative.
Experimental setup
Datasets: MovieGenVideoBench (extended prompts, calibration) VBench prompts (946 extended, evaluation)
Benchmarks: VBench
Metrics: FVD-FP LPIPS-FP VBench subject consistency background consistency motion smoothness aesthetic quality imaging quality coefficient of variation bitwidth-order agreement discriminability score
Baselines: MinMax AdaRound SmoothQuant PTQ4DiT SVDQuant
Builds on: AdaRound SmoothQuant PTQ4DiT SVDQuant Self-Forcing
Closest papers by meaning
- ARLON: Boosting Diffusion Transformers with Autoregressive Models for Long Video Generation (ICLR 2025 · similarity 0.89)
- Q-VDiT: Towards Accurate Quantization and Distillation of Video-Generation Diffusion Transformers (ICML 2025 · similarity 0.89)
- Q-DiT: Accurate Post-Training Quantization for Diffusion Transformers (CVPR 2025 · similarity 0.88)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.88)
- $\text{S}^2$Q-VDiT: Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation (NeurIPS 2025 · similarity 0.88)
- TCAQ-DM: Timestep-Channel Adaptive Quantization for Diffusion Models (AAAI 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.